animal trajectories by habitat use - A HITCHHIKER'S GUIDE TO WATERBIRD MIGRATION

M

ARIËLLE

L.

VAN

T

OOR

S

COTT

H. N

EWMAN

J

OHN

Y. T

AKEKAWA

M

ARTIN

W

EGMANN

K

AMRAN

S

AFI

Published inEcosphere,7(10), e01498

A

BSTRACT

Most animals live in seasonal environments, and experience very different conditions throughout the year. Behavioral strategies like migration, hi-bernation, and a life cycle adapted to the local seasonality help to cope with fluctuations in environmental conditions. Thus, how an individual utilizes the environment depends on both the current availability of habitat and the behavioral prerequisites of the individual at that time. While the in-creasing availability and richness of animal movement data has facilitated the development of algorithms that classify behavior by movement geom-etry, changes in the environmental correlates of animal movement have so far not been exploited for a behavioral annotation. Here, we suggest a method that uses these changes in individual-environment associations to divide animal location data into segments of higher ecological coherence, which we term niche segmentation. We use time series of random forest models to evaluate the transferability of habitat use over time to cluster observational data accordingly. We show that our method is able to iden-tify relevant changes in habitat use corresponding to both changes in the availability of habitat as well as how it was used using simulated data, and apply our method to a tracking dataset of common teal. The niche segmen-tation proved to be robust, and segmented models outperformed models neglecting the temporal dynamics of habitat use. Overall, we show that it is possible to classify animal trajectories based on changes of habitat use, similarly to geometric segmentation algorithms. We conclude that such an environmentally informed classification of animal trajectories can provide new insights into an individuals’ behavior, and enables us to make sensi-ble predictions of how suitasensi-ble areas might be connected by movement in space and time.

Keywords: Anas crecca, animal movement, common teal, habitat use, life history, migration, niche dynamics, random forest models, segmentation, simulation, species distribution model, transferability

Introduction

The technological advances that allow us to follow animals in the wild have revolutionized the field of movement ecology (Cagnacci et al., 2010; Hussey et al., 2015; Kays et al., 2015). Since the invention of simple tags such as bird bands a centennial ago, the miniaturization and increased efficiency in power consumption have given rise to modern tags which transmit or record locations of the tagged animals at an unprecedented rate. Animal location data have become ever more accurate in space and time, and the duration over which a single individual can be observed steadily increased. Thus, animal movement data have become not only more accurate, but also much more abundant. With the increased spatio-temporal resolution owed to the technological developments, the amount of details that can be gleaned from movement data has also increased. It is now possible to not only know where animals were and how much space they used, but also what they were doing during the time of observation. The contextualization of locations, i.e. the ability to put locations in a behavioral context, allows us to address important questions like how an individual allocates its time and energy to specific behaviors. Contextualization also is the basis on which we can associate resource distribution with a more detailed perspective of space use, as well as study the interactions between tagged individuals or even species, particularly in predator-prey dyads. The identification of behavior from animal trajectories thus provides a unique and important perspective on ecology in high detail in the wild.

While in the past mainly expert opinion was used for the behavioral classification of animal trajectories, the exponential growth of collected movement data as well as the necessity for repro-ducibility poses logistical limits on expert-based contextualization. Hence, behavioral classification is being increasingly based on algorithms often referred to as segmentation algorithms. These al-gorithms subset a behaviorally heterogeneous trajectory into a discrete number of segments that characterize distinct patterns representing coherent behavior (Gurarie et al., 2016). Current seg-mentation algorithms often rely on metrics such as speed and tortuousity of the trajectory, and are in general based on the geometry of the movement alone (e.g., Gurarie et al., 2009; Garriga et al., 2016).

Trajectories can, however, be characterized not only by their geometry, but also by the environ-mental conditions an individual was observed in. Certain behaviors like foraging and resting are often tied to a specific habitat, as observed in e.g. Spanish stone martens who use pastures for for-aging and orchards for resting (Santos and Santos-Reis, 2009). Similarly, the association between an animal and its environment can change with changing life history stages. For example, some species of migratory birds use very different habitat in their temperate breeding grounds compared to what they use in the tropical wintering areas (Nakazawa et al., 2004; Martínez-Meyer et al., 2004;

Batalden et al., 2007). Finally, habitat segregation can occur for different age classes of the same species, e.g. in cave salamanders (Ficetola et al., 2012). Thus, behavior across multiple temporal scales can be linked to the environment an individual is observed in. In the reverse conclusion, changes in the association between an individual and its environment could thus indicate changes in behavior. These changes in the association of individuals with their surrounding environment can be the consequence of different processes: the individuals moved to a different habitat, the en-vironment itself changed over time, or both happened in parallel, all of which can be indicative of a behavioral change.

We argue that the changes in the relationship between an individual and its environment can, similar to a segmentation based on movement geometry, be utilized for a behavioral segmentation of trajectories. In this study, we suggest a new class of segmentation algorithms that uses envi-ronmental correlates of a trajectory, rather than movement geometry, for behavioral classification.

We will utilize changes in the realized ecological niche of tracked individuals, and by comparing snapshots of their habitat use, introduce niche segmentation for movement trajectories. Besides being important for a better understanding of how individuals change their behavior in relation to changes in the environment, niche segmentation is also key to finding a minimum adequate

num-27

ber of time-explicit niche models when modelling habitat use of animals that undergo niche shifts in their life cycle. The identification of distinct realized niche volumes is pivotal to accurately pre-dict where animals will eventually be for a specific life history stage, or under specific environmental conditions.

Segmentation of animal trajectories by changes in habitat use

Our approach to niche segmentation is based on the classic habitat suitability, or ecological niche model (also termed species distribution models, SDM) and uses a measure of transferability to clus-ter subsets of a trajectory into segments based on their similarity in habitat use. Species distribu-tion models are derived from the Grinnellian and Hutchinsonian niche concepts (Grinnell, 1917;

Hutchinson, 1957; Soberón, 2007) and capture the environment species preferentially occur in to understand and approximate the potential distribution and abundance of individuals, populations, and species in space (e.g., Sattler et al., 2007). While originally developed for the estimation of a species’ ecological niche, SDMs are now also being used on the level of animal groups or individu-als to identify intra-species variation in habitat use, e.g. between age classes (Ficetola et al., 2012) or sexes (van Toor et al., 2011). One of the main advantages of SDMs is that they can describe a complex environment in terms of a single variable - habitat suitability - which is based on where a species or an individual occurs, and which environmental conditions are available. In contrast to a multivariate description of the environmental conditions at a given location, habitat suitability at a specific location can be easily compared across e.g. individuals or for different points in time.

In order to evaluate changes in habitat use along a series of animal locations, we divide an en-vironmentally annotated trajectory into non-overlapping windows of equal window sizeS(Figure 1, step 1). We model habitat use within each of these windows using an SDM, and subsequently compare habitat use between windows. To do so, we estimate the pair-wise similarity in habitat use between all windows using a measure of transferability, which we refer to as the Discriminatory Index (DI, see Appendix A). The DI quantifies how well a SDM can discriminate between presences and pseudo-absences, taking values between 1 (perfectly correct discrimination) and -1 (complete opposite prediction). By calculating the DI of all pair-wise comparisons of windows, we obtain a ma-trix of transferability that estimates similarity of habitat use across the SDMs of all windows (Figure 1, step 2). Finally, we group the windows into niche segments based on the similarity of habitat use through a clustering of the transferability matrix (Figure 1, step 3). To achieve this, we ordinate the transferability matrix such that windows for which DI is higher become placed closer to each other, and windows for which DI is lower become placed further apart in the two-dimensional space. A clustering is then applied to the ordination axis derived from the transferability matrix. In contrast to assuming a priori a number of clusters, we determine the number of clusters that produces the most compliant clustering using the respective cluster silhouettes (Rousseeuw, 1987). The resulting clustering of windows is finally used to annotate the original data with an environmentally informed segmentation, which we refer to as niche segmentation (Figure 1, step 4).

Testing the niche segmentation with simulated data

We test the our niche segmentation concept first on simulated environmental and movement data.

This allows us to evaluate whether known changes in habitat use can be detected using our ap-proach, an assessment which would be impossible to make in empirical animal trajectories. We consider the two different processes that can lead to shifts in the relationship between an individual and its environment, and their combination: i) changes in the environment available to the individ-uals without changes in habitat preference (niche following), ii) changes in the habitat preference of individuals without changes in the environment (niche switching), or iii) both processes in parallel (simultaneous change). By integrating all three of these processes into the simulated data, we can investigate whether changes in habitat preference or changes in the surrounding environment are more likely to be detected by the niche segmentation.

Figure 1 Flow chart outlining the process of the niche segmentation.

For our simulations we simulate movement trajectories using correlated random walks (CRW, e.g. Codling et al., 2008) which we biased by modelled preferences for the surrounding

environmen-29

tal conditions. Different from an unbiased CRW, the environmental conditions at every possible lo-cation are taken into account in an iterative step-wise simulation of the trajectories. This preference for a certain environmental condition is thus incorporated in the movement trajectory, and results in a realized niche reflecting this preference. We introduce changes in habitat use by switching this preference at specific known points in time mimicking niche switching, and/or gradually changing the environmental conditions to simulate niche following and the simultaneous change of both. We then use the niche segmentation concept to identify break points in the observed habitat use of the simulated trajectories, and test the ability of the method to reconstruct the underlying, simulated process.

Application to an empirical tracking dataset of common teal

Finally, we apply our niche segmentation approach to empirical tracking data of the common teal, a small species of dabbling duck with a wide distribution in the Northern hemisphere (Anas crecca, L.

1758). We make no a priori assumption about the timing of changes in habitat use over the annual cycle of the common teal, but instead reconstruct it from the data using the niche segmentation.

While this species is resident in some parts of its range, most of the populations are considered migratory. We therefore we expect to find changes in the temporal signature of habitat use at least as an effect of the migration between breeding and wintering ranges. Typical for empirical animal movement datasets is that they contain many idiosyncrasies making the analysis more challenging compared to simulated trajectories. We show how to address typically occurring characteristics of empirical data in the segmentation approach, such as irregular sampling or the inclusion of location error.

In the present study, we focus our main interest on seasonal changes of habitat use. We re-peat the niche segmentation using four different window sizesSto investigate how the choice ofS influences the resulting segmentation. Subsequently, we evaluate whether the niche segments de-tected by the segmentation reflect relevant changes in habitat use known for the species. We derive a set of SDMs specific to the niche segments identified by the segmentation to compare the spatial predictions of habitat suitability to the life-history of common teal as published in the literature.

Furthermore, we test how migration and catching site of individuals contributed to the resulting segmentation of the dataset. Finally, we compare the performance of SDMs derived for niche seg-ments to SDMs based on the complete dataset.

Materials and Methods Data preparation

Simulated data We used simulations to test the capability of our segmentation method to detect changes in habitat use under three different scenarios: i) niche following (constant habitat pref-erence in a changing environment), ii) niche switching (changing habitat prefpref-erence in a constant environment), and iii) a simultaneous change (changing habitat preference and changing environ-ment). We also allowed other parameters to vary, namely the number of niche segments introduced to the simulated data, and the size of windows the data were partitioned by (S). For each scenario, we varied all these parameters to estimate their effect on the accuracy of the method (Table 1), and computed 1000 replicates per scenario and number of niche segments whileS was chosen ran-domly. To simulate the movements of individuals in artificial landscapes, we used random fields (Schlather et al., 2015, R-packageRandomFields, version 3.1.8) and correlated random walks (CRWs) biased by habitat preference. We created landscapes using a Whittle-Matern covariance model on a grid of 250 by 250 cells of arbitrary size. This size was chosen because it was sufficiently large such that no simulated individual ever encountered an edge, thus avoiding edge effects. The habi-tat preference of individuals was sampled from the range between the 5%- and 95%-quantiles of

the available environmental conditions. We calculated a bias layer from the environment by nor-malizing the absolute difference between the individuals’ preference and the environmental value for each grid cell. For each of the simulations, we used a group of five individuals. The starting locations for the individuals were sampled from a circle around the grid center with a radius of 50 cells, and weighted by the corresponding bias layer values. For the biased CRWs, we created a two-dimensional kernel density of step lengths (Weibull distribution withk =2 andλ=1) and turning angles (wrapped Cauchy distribution withµ=0 rad andγ=0.9) with 100 by 100 grid cells, repre-senting 1000 potential steps of varying probability.

Table 1 Set-up for the different simulation experiments. Listed are all parameters used and how they were handled for each of the different scenarios. We computed 1000 replicates for every combination of param-eters, except for window size which was chosen randomly.

Parameter Niche following Niche switching Simultaneous change

Total number of replicates 4000 4000 4000

Habitat preference constant variable variable

Environment variable constant variable

Number of niche segments variable (3-5) variable (3-5) variable (3-5) Window sizeS random (50-500) random (50-500) random (50-500)

At each point in time (t), we calculated the putative end locations of all 1000 steps relative to the individual’s previous positionx_t₋₁. The realized step with the new position x_t was sampled from all possible steps weighted by the product of their probability (the kernel density) and the environ-mental bias of the corresponding locations. For the three different simulation scenarios, the data for the segmentation were prepared as follows: i) Niche following; We simulated a gradual change in the environment by shifting the values of the initial environmental layer by an arbitrary amount, and interpolated the number of desired segments between these two layers. Then, all individuals were allowed to take 100 steps on each of the layers, with the starting location on each layer corre-sponding to the last position on the previous layer. ii) Niche switching; For the each niche segment, all individuals were allowed to take 100 steps, resulting in a total of 500 presence locations for all 5 individuals. After reaching the last location, a new environmental preference was sampled and a new bias layer computed, and the process repeated. iii) Simultaneous change; In this case, we first created the environmental time series as in i), then sampled the number of species segments (ei-ther smaller or equal to the number of environmental segments). Again, for each of the segments, the individuals were allowed to take 100 steps in the corresponding environment and biased by the corresponding preference. We sampled 500 locations per segment as pseudo-absences to achieve a 1:1-ratio of presences vs. absences. Both presences and pseudo-absences were subsequently anno-tated with the environmental information in space and time. To prepare the data for segmentation, the complete dataset with information on position, presence or absence, environmental conditions, and the true niche segment for both environment and the individuals’ habitat preference was par-titioned into windows with the predefined sizeS. We provide a commented R-script that provides all necessary details to repeat the simulations in the Supporting Information (Data S1).

Tracking data of common teal

In addition to the simulation study, we tested our method on a tracking dataset of the common teal. Between 2007 and 2010, 34 individuals of common teal were caught at five different study sites (China, India, Kazakhstan, Egypt, and Turkey) and equipped with ARGOS tags before release (PTT-100; Microwave Telemetry, Columbia, Maryland, USA). These tracking data are part of a broader disease and migration ecology study implemented by the FAO and USGS. Locations were taken throughout the day, and 6448 positions for 22 individuals were obtained in total (Table 2). The me-dian sampling frequency for all individuals was 0.83 fixes per hour (25%-quantile: 0.24 fixes per

Table 2 A summary over the catching sites and corresponding sample sizes. The number of tracking days and locations are listed as a mean per individual.

Catching site Year Number of First fix taken tracking days locations individuals [mean±s.d] [mean±s.d.]

China/Lake Poyang 2007 3 Mar 18–20 336±243 289±157

Egypt 2009 8 Jan 18 – Nov 26 161±126 182±293

India 2008 5 Dec 09–18 133±81 211±112

Kazakhstan 2008 2 Sep 15–17 89±47 96±50

Turkey 2010 4 Feb 10–15 232±99 720±383

hour, 75%-quantile: 2.40 fixes per hour). These tracking data show characteristics that are typical of empirical data, and we addressed these using different approaches as described below:

Variation in location error A common confounder of tracking data is the use of different tags or geo-positioning sensors for the collection of animal movement data. The effect is a variation of location error in the dataset, and consequently the corresponding environmental conditions.

To account for this, we corrected for location error by using temporally explicit estimates of the individuals’ space use rather than the actual locations. From these utilization distributions (UD), we derived pseudo-presences that better reflect the actual distribution of individuals in space. We used the dynamic Brownian Bridge movement model (dBBMM, Kranstauber et al., 2012, R-package move, version 1.2.475) which estimates the UD of an individual from its movement path while also accounting for temporal autocorrelation and the spatial error of locations. Moreover, the dBBMM is time explicit, allowing us to estimate an individual’s UD at any given point in time. We applied the

Im Dokument A HITCHHIKER'S GUIDE TO WATERBIRD MIGRATION (Seite 23-45)